Create gradio_app.py
Browse files- gradio_app.py +118 -0
gradio_app.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import imageio
|
3 |
+
import os
|
4 |
+
import gradio as gr
|
5 |
+
from diffusers.schedulers import EulerAncestralDiscreteScheduler
|
6 |
+
from transformers import T5EncoderModel, T5Tokenizer
|
7 |
+
from allegro.pipelines.pipeline_allegro import AllegroPipeline
|
8 |
+
from allegro.models.vae.vae_allegro import AllegroAutoencoderKL3D
|
9 |
+
from allegro.models.transformers.transformer_3d_allegro import AllegroTransformer3DModel
|
10 |
+
|
11 |
+
import huggingface_hub
|
12 |
+
|
13 |
+
weights_dir = './allegro_weights'
|
14 |
+
os.makedirs(weights_dir, exist_ok=True)
|
15 |
+
|
16 |
+
huggingface_hub.snapshot_download(
|
17 |
+
repo_id='rhymes-ai/Allegro',
|
18 |
+
allow_patterns=[
|
19 |
+
'scheduler/**',
|
20 |
+
'text_encoder/**',
|
21 |
+
'tokenizer/**',
|
22 |
+
'transformer/**',
|
23 |
+
'vae/**',
|
24 |
+
],
|
25 |
+
local_dir=weights_dir,
|
26 |
+
local_dir_use_symlinks=False,
|
27 |
+
)
|
28 |
+
|
29 |
+
|
30 |
+
def single_inference(user_prompt, save_path, guidance_scale, num_sampling_steps, seed, enable_cpu_offload):
|
31 |
+
dtype = torch.bfloat16
|
32 |
+
|
33 |
+
# Load models
|
34 |
+
vae = AllegroAutoencoderKL3D.from_pretrained(
|
35 |
+
"weights_dir/vae/",
|
36 |
+
torch_dtype=torch.float32
|
37 |
+
).cuda()
|
38 |
+
vae.eval()
|
39 |
+
|
40 |
+
text_encoder = T5EncoderModel.from_pretrained("weights_dir/text_encoder/", torch_dtype=dtype)
|
41 |
+
text_encoder.eval()
|
42 |
+
|
43 |
+
tokenizer = T5Tokenizer.from_pretrained("weights_dir/tokenizer/")
|
44 |
+
|
45 |
+
scheduler = EulerAncestralDiscreteScheduler()
|
46 |
+
|
47 |
+
transformer = AllegroTransformer3DModel.from_pretrained("weights_dir/transformer/", torch_dtype=dtype).cuda()
|
48 |
+
transformer.eval()
|
49 |
+
|
50 |
+
allegro_pipeline = AllegroPipeline(
|
51 |
+
vae=vae,
|
52 |
+
text_encoder=text_encoder,
|
53 |
+
tokenizer=tokenizer,
|
54 |
+
scheduler=scheduler,
|
55 |
+
transformer=transformer
|
56 |
+
).to("cuda:0")
|
57 |
+
|
58 |
+
positive_prompt = """
|
59 |
+
(masterpiece), (best quality), (ultra-detailed), (unwatermarked),
|
60 |
+
{}
|
61 |
+
emotional, harmonious, vignette, 4k epic detailed, shot on kodak, 35mm photo,
|
62 |
+
sharp focus, high budget, cinemascope, moody, epic, gorgeous
|
63 |
+
"""
|
64 |
+
|
65 |
+
negative_prompt = """
|
66 |
+
nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality,
|
67 |
+
low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry.
|
68 |
+
"""
|
69 |
+
|
70 |
+
# Process user prompt
|
71 |
+
user_prompt = positive_prompt.format(user_prompt.lower().strip())
|
72 |
+
|
73 |
+
if enable_cpu_offload:
|
74 |
+
allegro_pipeline.enable_sequential_cpu_offload()
|
75 |
+
|
76 |
+
out_video = allegro_pipeline(
|
77 |
+
user_prompt,
|
78 |
+
negative_prompt=negative_prompt,
|
79 |
+
num_frames=88,
|
80 |
+
height=720,
|
81 |
+
width=1280,
|
82 |
+
num_inference_steps=num_sampling_steps,
|
83 |
+
guidance_scale=guidance_scale,
|
84 |
+
max_sequence_length=512,
|
85 |
+
generator=torch.Generator(device="cuda:0").manual_seed(seed)
|
86 |
+
).video[0]
|
87 |
+
|
88 |
+
# Save video
|
89 |
+
os.makedirs(os.path.dirname(save_path), exist_ok=True)
|
90 |
+
imageio.mimwrite(save_path, out_video, fps=15, quality=8)
|
91 |
+
|
92 |
+
return save_path
|
93 |
+
|
94 |
+
|
95 |
+
# Gradio interface function
|
96 |
+
def run_inference(user_prompt, guidance_scale, num_sampling_steps, seed, enable_cpu_offload):
|
97 |
+
save_path = "./output_videos/generated_video.mp4"
|
98 |
+
result_path = single_inference(user_prompt, save_path, guidance_scale, num_sampling_steps, seed, enable_cpu_offload)
|
99 |
+
return result_path
|
100 |
+
|
101 |
+
|
102 |
+
# Create Gradio interface
|
103 |
+
iface = gr.Interface(
|
104 |
+
fn=run_inference,
|
105 |
+
inputs=[
|
106 |
+
gr.Textbox(label="User Prompt"),
|
107 |
+
gr.Slider(minimum=0, maximum=20, step=0.1, label="Guidance Scale", value=7.5),
|
108 |
+
gr.Slider(minimum=10, maximum=200, step=1, label="Number of Sampling Steps", value=100),
|
109 |
+
gr.Slider(minimum=0, maximum=10000, step=1, label="Random Seed", value=42),
|
110 |
+
gr.Checkbox(label="Enable CPU Offload", value=False),
|
111 |
+
],
|
112 |
+
outputs=gr.Video(label="Generated Video"),
|
113 |
+
title="Allegro Video Generation",
|
114 |
+
description="Generate a video based on a text prompt using the Allegro pipeline."
|
115 |
+
)
|
116 |
+
|
117 |
+
# Launch the interface
|
118 |
+
iface.launch()
|